Goal, Swarm and Plan

A normal message starts one turn: the agent works until it answers you. Three modes in the composer's + menu change how the agent approaches the job:

• Goal: "Implement it, prove it, have it reviewed". The message becomes the goal of a verified change that runs on its own.
• Swarm: "Split the work across agents". The agent hands parts of the job to sub-agents that work at the same time.
• Plan mode: "Explore, ask, then plan before touching anything". For real work, the agent studies the task and asks for your approval before it changes anything.

Pick one from + and its icon appears beside the permission chip. It stays on for the next messages until you click the icon to remove it. Goal cannot be combined with Swarm or Plan mode; picking one turns the other off.

Goal mode

Goal mode is for a change you want made and checked from start to end. What you type is the goal. AgenC Core then runs a fixed pipeline in the background and reports each stage in the chat:

1. Intake: reads the goal and freezes it into a specification.
2. Worktree: takes an isolated Git worktree, so your own checkout stays untouched.
3. Plan: reads the repository and writes the plan the run is held to.
4. Implement: makes the change in the worktree, following the plan.
5. Verify: runs the verification command, and a verify agent checks the result.
6. Review: an independent reviewer reads the patch. It sees only the goal, the diff and the evidence, not the working agent's conversation.
7. Finalize: commits the change and seals the record of the evidence.

While it runs, the chat's ticker shows what the current stage is doing, and the final message of each stage is written into the chat. When the run succeeds, the chat says "Done: the change was implemented, proven by the verification command, and passed an independent review." When it fails or is stopped, it says how the run ended and "Nothing was merged."

A successful run leaves its result as a commit. Core pins it at the Git reference refs/agenc/runs/ followed by the run id, and never changes your own checkout or branch. Review the commit and merge it yourself.

Start a goal

1. Open a project, or stay on the welcome screen with a project chosen.
2. Click +, then Goal.
3. Describe the end state you want, in words a check could prove. "Every test in test/auth passes" is a goal. "Improve the auth code" is not.
4. Send.

Sent from the welcome screen, the goal opens a new session named after it. Sent from an open chat, the run reports in that chat.

The verification command

Core refuses a run without a command that proves the change, so the app picks the project's own. It looks in the project folder, in this order, and uses the first it finds:

• package.json scripts: npm test, then npm run build, check, lint or typecheck;
• a Makefile: make;
• Cargo.toml: cargo test;
• go.mod: go test ./...;
• pyproject.toml or pytest.ini: pytest.

If the project states none of these, the run still goes on a check that always passes, and the independent review is what holds the work to account. The ticker names the command while it runs.

Git is required

A goal runs in a Git worktree. If the folder is not a Git repository yet, the app initializes one (a main branch and an empty first commit) and says "Initialized a git repository in (folder) for the run." If that fails, the goal does not start and the app says why. A repository that contains the AgenC data folder cannot host a run; pick the project's own folder instead.

Spend and time limits

Every goal started from the app carries two limits, set in Settings > General > Goals:

• Spend limit: "A goal stops when it has cost this much." In US dollars, from 0 to 1,000. The default is 20.
• Time limit: "A goal stops when it has run this long." In hours, from 0 to 24 in half-hour steps. The default is 3.

Set either to 0 to remove that limit. Core enforces both. The defaults are meant to stay out of the way of a healthy run and to stop one that has stopped making progress.

Permissions during a goal

A goal runs in the session's permission mode. In "Ask for every tool" the run uses the pipeline's own default instead, "Approve edits automatically", because its agents work in the background where nobody could approve each edit. See Permissions and approvals.

The agents of a goal run are sub-agents, so they follow the effort and speed limit set for their provider, which by default is the lowest effort at standard speed. Raise it on the provider's card in Settings > Providers; see Sub-agents.

Stop, reload and restart

The stop button in the composer cancels the run; the chat says "Cancelling the goal run." A run belongs to Core, not to the window: if you reload or restart the app, it finds the run again, says "Goal run in progress." and keeps reporting until it ends.

Swarm

Swarm is for work that splits into parts that do not depend on each other, such as reviewing several modules or researching several options. With Swarm on, the agent spawns one named sub-agent per part, runs them at the same time, waits for them, and merges their results into one answer. For a small request it just does the work itself.

While sub-agents run, the ticker says how many are working, for example "3 agents working". Click it to open the agents panel, which lists each sub-agent with its task, time, tokens and tool uses, and lets you read its prompt and output. See Sub-agents.

Plan mode

Plan mode is for work you want to see planned before anything changes. When a message needs real work, the agent enters planning: it explores the project without editing, asks you what it needs to know, then presents a plan and waits. You approve it with Yes, implement this plan, or send it back with No, keep planning. For a small or obvious change, the agent just makes it.

Starting a message with the word "plan" (or "planning", "planea", "planifica") turns Plan mode on for you. A plan word elsewhere in the sentence does not.

Plan mode does not change the session's permission mode. When the plan is approved, the agent continues in the mode you had. To make a session strictly read-only instead, use the "Plan without changes" permission mode; see Permissions and approvals.

The task list

For longer work the agent often keeps a task list. It appears above the composer as "2 out of 5 tasks completed", with finished tasks struck through. Click its header to show or hide the tasks. The card goes away when the last task is done.